【作者】 段录平；

【导师】 周丽娟；

【作者基本信息】 哈尔滨理工大学 ， 计算机应用， 2007， 硕士

【摘要】 随着数据库技术的成熟应用和Internet的迅速发展,人们利用信息技术生产和搜集数据的能力大幅度提高,使得从大量数据中挖掘出有用的信息或知识成为一个迫切需要解决的问题。正是这种需求推动了数据挖掘兴起和数据挖掘技术的发展。数据挖掘经常要面对一些有噪声、杂乱、非线性的数据,而神经网络具有良好的鲁棒性、自适应性、并行处理、分布存储和高度容错性等特点,因此神经网络非常适合用来解决数据挖掘的一些问题。本文简单阐述了数据挖掘和人工神经网络的基本理论。在分析数据挖掘各种技术的基础上,对神经网络方法在数据挖掘中的应用进行了研究分析,接着着重研究了基于RBF神经网络的分类数据挖掘方法。在梯度算法基础推导出一种增量式的学习算法,在训练过程中该算法可以自适应调整网络参数。然后在IRIS数据库上进行分类实验,仿真实验结果表明该算法性能较好。在对RBF神经网络训练算法深入研究的基础上,本文采用了两阶段学习策略来加速学习收敛;提出动静相结合的隐含层设计方法来构造出较优的隐含层结构;提出采用误差校正的思想来改进RBF网络输出精度,并给出了其实现算法。并对这些改进算法在UCI数据库上进行了实验和对比分析,实验结果表明改进后的算法其性能均有明显提高。基于对数据挖掘和神经网络技术的研究,开发了一个主要用作实验平台的集成了本文各种算法的数据挖掘系统。本论文研究的基于RBF神经网络的数据挖掘方法具有一定的理论深度和实用价值,尤其创新的学习算法可以为相关的科研工作提供有益的参考。更多还原

【Abstract】 With the wide application of databases and sharp development of Internet, the capacity of utilizing information technology to manufacture and collect data has improved greatly. It is an urgent problem to mine useful information or knowledge from large databases or data warehouses. Therefore, data mining technology is developed rapidly to meet the need. But data mining (DM) often faces so much data which is noisy, disorder and nonlinear. Fortunately, artificial neural network (ANN) is suitable to solve the before-mentioned problems of DM because ANN has such merits as good robustness, adaptability, parallel-disposal, distributing-memory and high tolerating-error.This paper simply expounds the basic theory of DM and ANN. Based on the analysis of all kinds of data mining technology, gives a detailed discussion about the application of ANN method used in DM, and especially lays stress on the classification of DM which is based on radial basis function (RBF) neural networks. An incremental learning algorithm (ILA) is deduced from the gradient descend algorithm. ILA can adjust parameters of RBF networks adaptively driven by minimizing the error cost. And then applied it to resolve the IRIS problem, the experiment results show the algorithm has excellent performance.Based on the thorough research of the learning algorithm of RBF neural networks, a two-stage (TS) learning strategy is used to accelerate the convergence rate of traditional gradient descent algorithm; And a new method is proposed to design the hidden layer of RBF network by dynamic and static style; In order to improve the output precision of RBF networks, a creative algorithm named error-correcting (EC) is proposed for the first time in this paper. Afterward, these improving algorithms are tested by two UCI databases to evaluate their capability. Experiment results show that their performances are improved obviously.Additionally, basing on our research of data mining and neural networks technology, a data mining system for classification and prediction, which integrate all creative algorithm proposed in this paper, is developed primarily as an experiment platform.To some degree, this article has some theory significance and practical value. Especially some creative algorithm presented in this paper can provide helpful reference to related researcher.更多还原